AI search ranking factors: what actually gets you recommended

Before any of the AI-specific advice applies, a page has to enter the candidate set. In practice that means the traditional signals still govern the first gate: the page must be crawlable, indexed, reasonably fast, and relevant enough to the query to be pulled in. Analyses of AI Overview citations keep finding that the great majority of cited pages already rank on page one.

This is the least fashionable part of the answer and the most consequential. Teams frequently restructure content for extractability while leaving pages that are blocked, duplicated, canonicalised away, or too slow to fetch reliably. None of the downstream work matters if retrieval never happens.

The corollary is that a technical audit is the correct first move rather than a formatting pass. If you have not established which pages are actually indexed and competitive, you are optimizing in the dark. That baseline is what a site and visibility audit produces.

Once several pages are retrievable, the system needs a passage it can use. The property that matters is self-containment: can this block of text be lifted and still make sense. A paragraph beginning with a pronoun referring to the previous section fails that test even if its content is superior.

This is why heading phrasing and answer placement recur in every credible piece of guidance on the subject. A heading that matches the question, followed immediately by a direct answer, produces a passage that can be lifted without repair. The same information delivered as a conclusion after several paragraphs of context cannot.

It also explains why hedged writing performs badly. A sentence that surveys three possibilities without committing to one offers nothing extractable. Committing to a specific answer, then qualifying it underneath, gives the model something to use and the reader something to trust.

Models weight claims that multiple independent sources agree on, which is sensible behaviour for a system that cannot verify facts directly. The practical effect is that a statement appearing only on your domain is treated as an assertion, while the same statement repeated across a review site, an industry publication, and a community thread is treated as established.

This is the factor least amenable to on-site work and the one that most often explains why a competitor with a visibly worse page is cited instead of you. Their claim is corroborated; yours is not. No amount of formatting closes that gap.

It is also why consistency across your own properties matters more than it used to. When your pricing page, your documentation, and a third-party listing disagree about what you offer, there is no coherent claim to corroborate, and the system will use whichever version it finds easiest to lift.

Search systems reason about entities rather than strings, and they need to establish who you are, what category you belong to, and which topics you are credibly associated with. When that association is weak, you are simply not considered for questions you could answer well.

The signals that build it are unglamorous and mostly consistency. The same organisation name, description, and details wherever you appear. Structured data that states the entity explicitly rather than leaving it to be inferred, which is covered in our guide to schema markup for AI search. Topical depth that demonstrates coverage of a subject rather than one isolated page.

Depth is the part teams underinvest in. One page about a topic reads as an attempt to rank; twelve interlinked pages covering the subject properly read as expertise. This is the actual argument for content clusters, and it is a stronger argument now than it was when the goal was only internal link equity.

Keyword density has no visible role. Systems matching on meaning rather than string frequency do not reward repetition, and heavy repetition tends to make writing worse and therefore less extractable. The same applies to exact-match phrasing forced into headings at the cost of readability.

Publishing volume alone shows no reliable relationship to citation. Thirty generic pages monthly can raise coverage while producing nothing citable, because none of them contains a claim worth lifting. This is the tension at the centre of choosing between AI SEO tools built for throughput versus those built for governed execution.

Files that providers have not committed to reading, llms.txt among them, do not currently influence AI search visibility despite frequent claims otherwise. We covered the evidence in our llms.txt guide. Publish one if it suits your documentation, but do not count it as a ranking factor.